PitchBook-新兴空间简报：LLM代理（英）

1Originally published September 18, 2024pbinstitutionalresearch@pitchbook.comAli Javaheri Analyst, Emerging Technologyali.javaheri@pitchbook.comEMERGING SPACE BRIEFLLM Agents1: Monte Carlo tree search is a decision-making algorithm in which a computer simulates many possible outcomes, learns from them, and then makes the best choice based on those results.Trending companiesOverviewLarge language model (LLM) agents represent an emerging technology in the field of artificial intelligence. These systems use advanced language models to process complex instructions, develop plans, and execute tasks across various domains. Key characteristics of LLM agents include task decomposition capabilities, integration with external tools and APIs, adaptive learning potential, and context retention across interactions.The development of LLM agents has potential implications for productivity enhancement and operational streamlining across industries.BackgroundThe evolution of LLM agents is rooted in advancements in natural language processing and machine learning. While agents are an old concept in reinforcement learning research, LLM agents have appeared more recently, developed in tandem with transformer models. The development of LLM agents has the potential to enhance productivity and streamline operations across industries.A few key milestones in the modern concept include:• Early 2010s: Deep learning techniques and neural networks advance, setting the stage for more sophisticated language models.• 2017: The Transformer architecture is introduced, significantly improving the efficiency of training LLMs. DeepMind’s agentic model, AlphaZero, learned to master games including Go, chess, and shogi (Japanese chess), achieving superhuman performance in all three games and showing the ability of AI to execute tasks based on reinforcement learning using Monte Carlo tree search.1• 2018-2020: OpenAI releases the GPT (Generative Pretrained Transformer) series, demonstrating enhanced language understanding and generation capabilities.• 2020-2022: Task-specific fine-tuning and prompt engineering techniques emerge, enabling more targeted applications of language models.LLM agents VC deal activity $92.7$490.0$323.1$1,317.9$1,036.219273810411120202021202220232024*Deal value ($M)Deal countSource: PitchBook • Geography: Global*As of September 17, 2024Note: Excluding OpenAIFor access to more of this data and PitchBook’s Emerging Spaces tool, access a free trial link here.2Emerging Space Brief: LLM Agents2: “The Brief History of AI Agents (2023-2024),” YouTube, uploaded by swyx, August 14, 2024. 3: Ibid. 4: “LLM Agents,” Prompt Engineering Guide, n.d., accessed September 17, 2024. 5: “Multi-Agent Conversation Framework,” GitHub, 2024, accessed September 17, 2024.• April 2023: Open-source LLM agent project Auto-GPT becomes the fastest-growing GitHub repository in history, showcasing the potential for autonomous AI agents to complete multistep tasks.2• July 2023: OpenAI releas